sakshi-pii-in

Token-classification model (MuRIL-base fine-tune) detecting Indian PII in unstructured text: PERSON, ADDRESS, DOB — including inside code-mixed Hinglish and Devanagari/Bengali/Tamil content.

Built by RotaVision as Layer 2 of the Sakshi agent-governance platform's ingest tokenization. Structured identifiers (Aadhaar, PAN, IFSC, mobile, account numbers) are handled by a deterministic regex+checksum layer upstream and are deliberately out of scope for this model — pair it with checksum detectors, don't ask it to find Aadhaar numbers.

Training data

100% synthetic: templated sentences, JSON records, and chat transcripts across English, Hindi (Devanagari), Hinglish, Bengali, and Tamil, populated from public-knowledge name/place gazetteers. No scraped data, no real persons, no customer data. Hard negatives include order/ticket/invoice identifiers and checksum-valid synthetic Aadhaar numbers (unlabeled — the deterministic layer owns those).

Evaluation (frozen 2,000-example set, disjoint names/places/templates)

Metric	Result
PERSON F1	0.988 (P 0.989 / R 0.988)
ADDRESS F1	0.956 (P 0.957 / R 0.959)
DOB F1	0.998
Micro F1 / recall	0.979 / 0.981
CPU batch latency (B=16)	p50 5.7 / p95 6.2 ms per record

Operating threshold: 0.25. Confidence scores are diffuse (correct spans score 0.28–0.6); calibration is a v1 item. Selection was recall-prioritized (F-beta, β=2): a missed identity is a liability, a false positive is a tokenized order id.

from transformers import pipeline

ner = pipeline("token-classification", model="rotalabs/sakshi-pii-in",
               aggregation_strategy="first")
spans = [s for s in ner("Priya Sharma, DOB 14/03/1991, Andheri West se hai")
         if s["score"] >= 0.25]

Limitations

• Five language settings in training; other Indic scripts (Malayalam, Telugu, Kannada, Odia…) are best-effort — expect lower recall.
• Synthetic-only training under-represents OCR noise and SMS shorthand.
• Finds PII in text; does not decide lawfulness of processing. Not a consent-management or data-discovery tool.
• Machine identifiers (fingerprints, digests, version strings) should be exempted from NER at the application layer — see the Sakshi source for the pattern.

Provenance

Base model google/muril-base-cased (Apache-2.0). Training data fully synthetic, generator and frozen eval set maintained in the Sakshi repository. Released by RotaVision under the rotalabs open-source commons, Apache-2.0.